Two Step Relevance Feedback for Semantic Disambiguation in Image Retrieval

This paper presents a new approach to the problem of feature weighting for content based image retrieval. If a query image admits to multiple interpretations, user feedback on the set of returned images can be an effective tool to improve retrieval performance in subsequent rounds. For this to work, however, the first results set has to include representatives of the semantic facet of interest. We will argue that relevance feedback techniques that fix the distance metric for the first retrieval round are semantically biased and may fail to distil relevant semantic facets thus limiting the scope of relevance feedback. Our approach is based on the notion of the NN k of a query image, defined as the set of images that are nearest neighbours of the query under some instantiation of a parametrised distance metric. Different neighbours may be viewed as representing different meanings of the query. By associating each NN k with the parameters for which it was ranked closest to the query, the selection of relevant NN k by a user provides us with parameters for the second retrieval round. We evaluate this two step relevance feedback technique on two collections and compare it to an alternative relevance feedback method and to an oracle for which the optimal parameter values are known.